Abstract: The existing music recommendation methods often use similarity or correlation to generate recommended music list, those methods don‘t consider the volatility of users  interest reflected by the historical music behavior, which influences the recommendation accuracy. To solve this problem, this paper proposes a music recommendation method based on Fisher linear discriminant analysis and volatility sequence. In the beginning, this method obtains the social tags and audio features of music to compute the projection direction which has the minimum within-class scatter and maximum between-class scatter, by using projection transformation and Fisher discriminant criterion. This projection direction is also the best direction of classification. Then it takes music type base point as center, type volatility range as radius to acquire users’ preferred music type, and based on which to generate the recommendation list. This paper presents the empirical experiments in a real data set LFM, the results show that the proposed method can achieve better P@R and coverage rate, which means it efficiently improves recommendation accuracy and quality.

Key words: Fisher linear discriminant analysis, volatility sequence, music type base point, social tags, music recommender systems

摘要：

现有的音乐推荐方法多是采用不同的历史偏好相关性度量方法直接为用户生成推荐音乐列表，而不考虑用户历史喜好音乐行为所体现出的用户兴趣的波动性，影响了推荐音乐的准确率。针对这个问题，提出了一种融合Fisher线性判别分析与波动序列的音乐行为偏好获取方法。首先获取音乐的社会化标签与音频特征，采用Fisher线性判别分析对两类样本数据进行特征融合，通过投影变换并引入Fisher判别准则，获取具有最大类间离散度，最小类内离散度的音乐特征分类方向。然后结合用户的历史喜好音乐获取音乐类型基点、类型波动幅度，再以音乐类型基点为中心，以类型波动幅度为半径获取用户的喜好音乐类型，并据此为用户生成推荐音乐列表。在真实数据集LFM上的仿真实验结果表明，所提出方法能够取得更好的P@R值与覆盖率，提升了音乐推荐精度与推荐质量。

关键词: Fisher线性判别分析, 波动序列, 音乐类型基点, 社会化标签, 音乐推荐系统